[HLSL] Implement the f16tof32() intrinsic (llvm#165860)

Implement the f16tof32() intrinsic, including DXILand SPIRV codegen, and
associated tests.

Fixes llvm#99112

---------

Co-authored-by: Tim Corringham <[email protected]>

Author: tcorringham

clang/include/clang/Basic/Builtins.td

@@ -5235,6 +5235,12 @@ def HLSLGetSpirvSpecConstant : LangBuiltin<"HLSL_LANG">, HLSLScalarTemplate {
   let Prototype = "T(unsigned int, T)";
 }
 
+def HLSLF16ToF32 : LangBuiltin<"HLSL_LANG"> {
+  let Spellings = ["__builtin_hlsl_elementwise_f16tof32"];
+  let Attributes = [NoThrow, Const, CustomTypeChecking];
+  let Prototype = "void(...)";
+}
+
 // Builtins for XRay.
 def XRayCustomEvent : Builtin {
   let Spellings = ["__xray_customevent"];

clang/lib/CodeGen/CGHLSLBuiltins.cpp

@@ -160,6 +160,57 @@ static Value *handleHlslSplitdouble(const CallExpr *E, CodeGenFunction *CGF) {
   return LastInst;
 }
 
+static Value *handleElementwiseF16ToF32(CodeGenFunction &CGF,
+                                        const CallExpr *E) {
+  Value *Op0 = CGF.EmitScalarExpr(E->getArg(0));
+  QualType Op0Ty = E->getArg(0)->getType();
+  llvm::Type *ResType = CGF.FloatTy;
+  uint64_t NumElements = 0;
+  if (Op0->getType()->isVectorTy()) {
+    NumElements =
+        E->getArg(0)->getType()->castAs<clang::VectorType>()->getNumElements();
+    ResType =
+        llvm::VectorType::get(ResType, ElementCount::getFixed(NumElements));
+  }
+  if (!Op0Ty->hasUnsignedIntegerRepresentation())
+    llvm_unreachable(
+        "f16tof32 operand must have an unsigned int representation");
+
+  if (CGF.CGM.getTriple().isDXIL())
+    return CGF.Builder.CreateIntrinsic(ResType, Intrinsic::dx_legacyf16tof32,
+                                       ArrayRef<Value *>{Op0}, nullptr,
+                                       "hlsl.f16tof32");
+
+  if (CGF.CGM.getTriple().isSPIRV()) {
+    // We use the SPIRV UnpackHalf2x16 operation to avoid the need for the
+    // Int16 and Float16 capabilities
+    auto UnpackType =
+        llvm::VectorType::get(CGF.FloatTy, ElementCount::getFixed(2));
+    if (NumElements == 0) {
+      // a scalar input - simply extract the first element of the unpacked
+      // vector
+      Value *Unpack = CGF.Builder.CreateIntrinsic(
+          UnpackType, Intrinsic::spv_unpackhalf2x16, ArrayRef<Value *>{Op0});
+      return CGF.Builder.CreateExtractElement(Unpack, (uint64_t)0);
+    } else {
+      // a vector input - build a congruent output vector by iterating through
+      // the input vector calling unpackhalf2x16 for each element
+      Value *Result = PoisonValue::get(ResType);
+      for (uint64_t i = 0; i < NumElements; i++) {
+        Value *InVal = CGF.Builder.CreateExtractElement(Op0, i);
+        Value *Unpack = CGF.Builder.CreateIntrinsic(
+            UnpackType, Intrinsic::spv_unpackhalf2x16,
+            ArrayRef<Value *>{InVal});
+        Value *Res = CGF.Builder.CreateExtractElement(Unpack, (uint64_t)0);
+        Result = CGF.Builder.CreateInsertElement(Result, Res, i);
+      }
+      return Result;
+    }
+  }
+
+  llvm_unreachable("Intrinsic F16ToF32 not supported by target architecture");
+}
+
 static Value *emitBufferStride(CodeGenFunction *CGF, const Expr *HandleExpr,
                                LValue &Stride) {
   // Figure out the stride of the buffer elements from the handle type.
@@ -579,6 +630,9 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned BuiltinID,
         /*ReturnType=*/X->getType(), CGM.getHLSLRuntime().getDegreesIntrinsic(),
         ArrayRef<Value *>{X}, nullptr, "hlsl.degrees");
   }
+  case Builtin::BI__builtin_hlsl_elementwise_f16tof32: {
+    return handleElementwiseF16ToF32(*this, E);
+  }
   case Builtin::BI__builtin_hlsl_elementwise_frac: {
     Value *Op0 = EmitScalarExpr(E->getArg(0));
     if (!E->getArg(0)->getType()->hasFloatingRepresentation())

clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h

@@ -1052,6 +1052,27 @@ float3 exp2(float3);
 _HLSL_BUILTIN_ALIAS(__builtin_elementwise_exp2)
 float4 exp2(float4);
 
+//===----------------------------------------------------------------------===//
+// f16tof32 builtins
+//===----------------------------------------------------------------------===//
+
+/// \fn float f16tof32(uint x)
+/// \brief Returns the half value stored in the low 16 bits of the uint arg
+/// converted to a float.
+/// \param x The uint containing two half values.
+///
+/// The float value of the half value found in the low 16 bits of the \a xi
+/// parameter.
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_f16tof32)
+float f16tof32(uint);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_f16tof32)
+float2 f16tof32(uint2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_f16tof32)
+float3 f16tof32(uint3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_f16tof32)
+float4 f16tof32(uint4);
+
 //===----------------------------------------------------------------------===//
 // firstbithigh builtins
 //===----------------------------------------------------------------------===//

clang/lib/Sema/SemaHLSL.cpp

@@ -2802,6 +2802,23 @@ static bool CheckUnsignedIntRepresentation(Sema *S, SourceLocation Loc,
   return false;
 }
 
+static bool CheckExpectedBitWidth(Sema *S, CallExpr *TheCall,
+                                  unsigned ArgOrdinal, unsigned Width) {
+  QualType ArgTy = TheCall->getArg(0)->getType();
+  if (auto *VTy = ArgTy->getAs<VectorType>())
+    ArgTy = VTy->getElementType();
+  // ensure arg type has expected bit width
+  uint64_t ElementBitCount =
+      S->getASTContext().getTypeSizeInChars(ArgTy).getQuantity() * 8;
+  if (ElementBitCount != Width) {
+    S->Diag(TheCall->getArg(0)->getBeginLoc(),
+            diag::err_integer_incorrect_bit_count)
+        << Width << ElementBitCount;
+    return true;
+  }
+  return false;
+}
+
 static void SetElementTypeAsReturnType(Sema *S, CallExpr *TheCall,
                                        QualType ReturnType) {
   auto *VecTyA = TheCall->getArg(0)->getType()->getAs<VectorType>();
@@ -2961,24 +2978,16 @@ bool SemaHLSL::CheckBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) {
                                    CheckUnsignedIntVecRepresentation))
       return true;
 
-    auto *VTy = TheCall->getArg(0)->getType()->getAs<VectorType>();
     // ensure arg integers are 32-bits
-    uint64_t ElementBitCount = getASTContext()
-                                   .getTypeSizeInChars(VTy->getElementType())
-                                   .getQuantity() *
-                               8;
-    if (ElementBitCount != 32) {
-      SemaRef.Diag(TheCall->getBeginLoc(),
-                   diag::err_integer_incorrect_bit_count)
-          << 32 << ElementBitCount;
+    if (CheckExpectedBitWidth(&SemaRef, TheCall, 0, 32))
       return true;
-    }
 
     // ensure both args are vectors of total bit size of a multiple of 64
+    auto *VTy = TheCall->getArg(0)->getType()->getAs<VectorType>();
     int NumElementsArg = VTy->getNumElements();
     if (NumElementsArg != 2 && NumElementsArg != 4) {
       SemaRef.Diag(TheCall->getBeginLoc(), diag::err_vector_incorrect_bit_count)
-          << 1 /*a multiple of*/ << 64 << NumElementsArg * ElementBitCount;
+          << 1 /*a multiple of*/ << 64 << NumElementsArg * 32;
       return true;
     }
 
@@ -3295,7 +3304,7 @@ bool SemaHLSL::CheckBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) {
     break;
   }
   // Note these are llvm builtins that we want to catch invalid intrinsic
-  // generation. Normal handling of these builitns will occur elsewhere.
+  // generation. Normal handling of these builtins will occur elsewhere.
   case Builtin::BI__builtin_elementwise_bitreverse: {
     // does not include a check for number of arguments
     // because that is done previously
@@ -3405,6 +3414,30 @@ bool SemaHLSL::CheckBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) {
     }
     break;
   }
+  case Builtin::BI__builtin_hlsl_elementwise_f16tof32: {
+    if (SemaRef.checkArgCount(TheCall, 1))
+      return true;
+    if (CheckAllArgTypesAreCorrect(&SemaRef, TheCall,
+                                   CheckUnsignedIntRepresentation))
+      return true;
+    // ensure arg integers are 32 bits
+    if (CheckExpectedBitWidth(&SemaRef, TheCall, 0, 32))
+      return true;
+    // check it wasn't a bool type
+    QualType ArgTy = TheCall->getArg(0)->getType();
+    if (auto *VTy = ArgTy->getAs<VectorType>())
+      ArgTy = VTy->getElementType();
+    if (ArgTy->isBooleanType()) {
+      SemaRef.Diag(TheCall->getArg(0)->getBeginLoc(),
+                   diag::err_builtin_invalid_arg_type)
+          << 1 << /* scalar or vector of */ 5 << /* unsigned int */ 3
+          << /* no fp */ 0 << TheCall->getArg(0)->getType();
+      return true;
+    }
+
+    SetElementTypeAsReturnType(&SemaRef, TheCall, getASTContext().FloatTy);
+    break;
+  }
   }
   return false;
 }

clang/test/CodeGenHLSL/builtins/f16tof32-builtin.hlsl

@@ -0,0 +1,30 @@
+// RUN: %clang_cc1 -std=hlsl202x -finclude-default-header -x hlsl -triple \
+// RUN:   dxil-pc-shadermodel6.3-library %s -emit-llvm -disable-llvm-passes \
+// RUN:   -o - | FileCheck %s
+
+// CHECK: define hidden noundef nofpclass(nan inf) float
+// CHECK: %hlsl.f16tof32 = call reassoc nnan ninf nsz arcp afn float @llvm.dx.legacyf16tof32.i32(i32 %0)
+// CHECK: ret float %hlsl.f16tof32
+// CHECK: declare float @llvm.dx.legacyf16tof32.i32(i32)
+float test_scalar(uint p0) { return __builtin_hlsl_elementwise_f16tof32(p0); }
+
+// CHECK: define hidden noundef nofpclass(nan inf) <2 x float>
+// CHECK: %hlsl.f16tof32 = call reassoc nnan ninf nsz arcp afn <2 x float> @llvm.dx.legacyf16tof32.v2i32(<2 x i32> %0)
+// CHECK: ret <2 x float> %hlsl.f16tof32
+// CHECK: declare <2 x float> @llvm.dx.legacyf16tof32.v2i32(<2 x i32>)
+float2 test_uint2(uint2 p0) { return __builtin_hlsl_elementwise_f16tof32(p0); }
+
+// CHECK: define hidden noundef nofpclass(nan inf) <3 x float> @_Z10test_uint3Dv3_j(<3 x i32> noundef %p0) #0 {
+// CHECK: %hlsl.f16tof32 = call reassoc nnan ninf nsz arcp afn <3 x float> @llvm.dx.legacyf16tof32.v3i32(<3 x i32> %0)
+// CHECK: ret <3 x float> %hlsl.f16tof32
+// CHECK: declare <3 x float> @llvm.dx.legacyf16tof32.v3i32(<3 x i32>)
+float3 test_uint3(uint3 p0) { return __builtin_hlsl_elementwise_f16tof32(p0); }
+
+// CHECK: define hidden noundef nofpclass(nan inf) <4 x float> @_Z10test_uint4Dv4_j(<4 x i32> noundef %p0) #0 {
+// CHECK: %hlsl.f16tof32 = call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.dx.legacyf16tof32.v4i32(<4 x i32> %0)
+// CHECK: ret <4 x float> %hlsl.f16tof32
+// CHECK: declare <4 x float> @llvm.dx.legacyf16tof32.v4i32(<4 x i32>)
+float4 test_uint4(uint4 p0) { return __builtin_hlsl_elementwise_f16tof32(p0); }
+
+
+

clang/test/CodeGenHLSL/builtins/f16tof32.hlsl

@@ -0,0 +1,30 @@
+// RUN: %clang_cc1 -std=hlsl202x -finclude-default-header -x hlsl -triple \
+// RUN:   dxil-pc-shadermodel6.3-library %s -emit-llvm -disable-llvm-passes \
+// RUN:   -o - | FileCheck %s
+
+// CHECK: define hidden noundef nofpclass(nan inf) float
+// CHECK: %hlsl.f16tof32 = call reassoc nnan ninf nsz arcp afn float @llvm.dx.legacyf16tof32.i32(i32 %0)
+// CHECK: ret float %hlsl.f16tof32
+// CHECK: declare float @llvm.dx.legacyf16tof32.i32(i32)
+float test_scalar(uint p0) { return f16tof32(p0); }
+
+// CHECK: define hidden noundef nofpclass(nan inf) <2 x float>
+// CHECK: %hlsl.f16tof32 = call reassoc nnan ninf nsz arcp afn <2 x float> @llvm.dx.legacyf16tof32.v2i32(<2 x i32> %0)
+// CHECK: ret <2 x float> %hlsl.f16tof32
+// CHECK: declare <2 x float> @llvm.dx.legacyf16tof32.v2i32(<2 x i32>)
+float2 test_uint2(uint2 p0) { return f16tof32(p0); }
+
+// CHECK: define hidden noundef nofpclass(nan inf) <3 x float> @_Z10test_uint3Dv3_j(<3 x i32> noundef %p0) #0 {
+// CHECK: %hlsl.f16tof32 = call reassoc nnan ninf nsz arcp afn <3 x float> @llvm.dx.legacyf16tof32.v3i32(<3 x i32> %0)
+// CHECK: ret <3 x float> %hlsl.f16tof32
+// CHECK: declare <3 x float> @llvm.dx.legacyf16tof32.v3i32(<3 x i32>)
+float3 test_uint3(uint3 p0) { return f16tof32(p0); }
+
+// CHECK: define hidden noundef nofpclass(nan inf) <4 x float> @_Z10test_uint4Dv4_j(<4 x i32> noundef %p0) #0 {
+// CHECK: %hlsl.f16tof32 = call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.dx.legacyf16tof32.v4i32(<4 x i32> %0)
+// CHECK: ret <4 x float> %hlsl.f16tof32
+// CHECK: declare <4 x float> @llvm.dx.legacyf16tof32.v4i32(<4 x i32>)
+float4 test_uint4(uint4 p0) { return f16tof32(p0); }
+
+
+

clang/test/SemaHLSL/BuiltIns/f16tof32-errors.hlsl

@@ -0,0 +1,134 @@
+// RUN: %clang_cc1 -finclude-default-header -x hlsl -triple dxil-pc-shadermodel6.6-library %s -fnative-int16-type -emit-llvm-only -disable-llvm-passes -verify
+
+float builtin_f16tof32_too_few_arg() {
+  return __builtin_hlsl_elementwise_f16tof32();
+  // expected-error@-1 {{too few arguments to function call, expected 1, have 0}}
+  // expected-note@hlsl/hlsl_alias_intrinsics.h:* 4 {{candidate function not viable: requires 1 argument, but 0 were provided}}
+}
+
+float builtin_f16tof32_too_many_arg(uint p0) {
+  return __builtin_hlsl_elementwise_f16tof32(p0, p0);
+  // expected-error@-1 {{too many arguments to function call, expected 1, have 2}}
+  // expected-note@hlsl/hlsl_alias_intrinsics.h:* 4 {{candidate function not viable: requires 1 argument, but 2 were provided}}
+}
+
+float builtin_f16tof32_bool(bool p0) {
+  return __builtin_hlsl_elementwise_f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'bool')}}
+}
+
+float builtin_f16tof32_bool4(bool4 p0) {
+  return __builtin_hlsl_elementwise_f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'bool4' (aka 'vector<bool, 4>')}}
+}
+
+float builtin_f16tof32_short(short p0) {
+  return __builtin_hlsl_elementwise_f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'short')}}
+}
+
+float builtin_f16tof32_unsigned_short(unsigned short p0) {
+  return __builtin_hlsl_elementwise_f16tof32(p0);
+  // expected-error@-1 {{incorrect number of bits in integer (expected 32 bits, have 16)}}
+}
+
+float builtin_f16tof32_int(int p0) {
+  return __builtin_hlsl_elementwise_f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'int')}}
+}
+
+float builtin_f16tof32_int64_t(long p0) {
+  return __builtin_hlsl_elementwise_f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'long')}}
+}
+
+float2 builtin_f16tof32_int2_to_float2_promotion(int2 p0) {
+  return __builtin_hlsl_elementwise_f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'int2' (aka 'vector<int, 2>'))}}
+}
+
+float builtin_f16tof32_half(half p0) {
+  return __builtin_hlsl_elementwise_f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'half')}}
+}
+
+float builtin_f16tof32_half4(half4 p0) {
+  return __builtin_hlsl_elementwise_f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'half4' (aka 'vector<half, 4>'))}}
+}
+
+float builtin_f16tof32_float(float p0) {
+  return __builtin_hlsl_elementwise_f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'float')}}
+}
+
+float builtin_f16tof32_double(double p0) {
+  return __builtin_hlsl_elementwise_f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'double')}}
+}
+
+float f16tof32_too_few_arg() {
+  return f16tof32();
+  // expected-error@-1 {{no matching function for call to 'f16tof32'}}
+}
+
+float f16tof32_too_many_arg(uint p0) {
+  return f16tof32(p0, p0);
+  // expected-error@-1 {{no matching function for call to 'f16tof32'}}
+}
+
+float f16tof32_bool(bool p0) {
+  return f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'bool')}}
+}
+
+float f16tof32_bool3(bool3 p0) {
+  return f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'bool3' (aka 'vector<bool, 3>'))}}
+}
+
+
+float f16tof32_int16_t(short p0) {
+  return f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'short')}}
+}
+
+float f16tof32_int16_t(unsigned short p0) {
+  return f16tof32(p0);
+  // expected-error@-1 {{incorrect number of bits in integer (expected 32 bits, have 16)}}
+}
+
+float f16tof32_int(int p0) {
+  return f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'int')}}
+}
+
+float f16tof32_int64_t(long p0) {
+  return f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'long')}}
+}
+
+float2 f16tof32_int2_to_float2_promotion(int3 p0) {
+  return f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'int3' (aka 'vector<int, 3>'))}}
+}
+
+float f16tof32_half(half p0) {
+  return f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'half')}}
+}
+
+float f16tof32_half2(half2 p0) {
+  return f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'half2' (aka 'vector<half, 2>'))}}
+}
+
+float f16tof32_float(float p0) {
+  return f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'float')}}
+}
+
+float f16tof32_double(double p0) {
+  return f16tof32(p0);
+  // expected-error@-1 {{1st argument must be a scalar or vector of unsigned integer types (was 'double')}}
+}